Extensive data engineering to the rescue: building a multi-species katydid detector from unbalanced, atypical...

Area of Science:

• Bioacoustics
• Ecology
• Machine Learning

Background:

• Passive acoustic monitoring (PAM) is valuable for ecosystem studies but faces challenges in tropical insect monitoring.
• Neotropical katydid calls are complex, species-specific, and difficult to distinguish using traditional PAM due to subtle variations and environmental noise.
• Limited and imbalanced training data further complicates automated species identification.

Purpose of the Study:

• To develop a deep learning-based solution for automated recognition of 31 katydid species in a biodiverse Panamanian forest.
• To address challenges posed by complex katydid vocalizations, low source levels, ambient noise, and limited training data.
• To create an open-source toolbox (Koogu) for advanced bioacoustic analysis.

Main Methods:

• Applied rigorous data engineering, including controlled playback re-recordings and physics-based data augmentation, to enhance input variance.
• Tuned signal-processing, model, and training parameters for a custom deep learning solution.
• Incorporated developed methods into Koogu, an open-source Python toolbox for bioacoustic analysis.

Main Results:

• Successfully developed a deep learning model for automated recognition of 31 katydid species.
• Overcame limitations of small, imbalanced, and domain-mismatched datasets through advanced data engineering and augmentation.
• The developed methods are integrated into the Koogu toolbox, offering a practical solution for bioacoustic analysis.

Conclusions:

• The study presents a robust deep learning approach for automated katydid species identification using PAM.
• The Koogu toolbox provides a valuable resource for enhancing insect biodiversity monitoring in tropical ecosystems.
• This work contributes to developing a global toolkit for insect biodiversity monitoring.